alibaba--mnn/tools/train/source/demo/MobilenetV2Utils.cpp

//
//  MobilenetV2Utils.cpp
//  MNN
//
//  Created by MNN on 2020/01/08.
//  Copyright © 2018, Alibaba Group Holding Limited
//

#include "MobilenetV2Utils.hpp"
#include <MNN/expr/Executor.hpp>
#include <MNN/expr/Optimizer.hpp>
#include <cmath>
#include <iostream>
#include <vector>
#include "DataLoader.hpp"
#include "DemoUnit.hpp"
#include "NN.hpp"
#include "SGD.hpp"
//#define MNN_OPEN_TIME_TRACE
#include <MNN/AutoTime.hpp>
#include "ADAM.hpp"
#include "LearningRateScheduler.hpp"
#include "Loss.hpp"
#include "RandomGenerator.hpp"
#include "Transformer.hpp"
#include "ImageDataset.hpp"
#include "module/PipelineModule.hpp"
#include "cpp/ConvertToFullQuant.hpp"

using namespace MNN;
using namespace MNN::Express;
using namespace MNN::Train;

void MobilenetV2Utils::train(MNNForwardType backend, int threadNumber, std::shared_ptr<Module> model, const int numClasses, const int addToLabel,
                                std::string trainImagesFolder, std::string trainImagesTxt,
                                std::string testImagesFolder, std::string testImagesTxt, const int quantBits, int size) {
    auto exe = Executor::getGlobalExecutor();
    BackendConfig config;
    exe->setGlobalExecutorConfig(backend, config, threadNumber);
    std::shared_ptr<SGD> solver(new ADAM(model));
    solver->setMomentum(0.9f);
    // solver->setMomentum2(0.99f);
    solver->setWeightDecay(0.00004f);

    auto converImagesToFormat  = CV::RGB;
    int resizeHeight           = size;
    int resizeWidth            = size;
    std::vector<float> means = {127.5f, 127.5f, 127.5f};
    std::vector<float> scales = {1/127.5f, 1/127.5f, 1/127.5f};
    std::vector<float> cropFraction = {0.875f, 0.875f}; // center crop fraction for height and width
    if (size == 32) {
        cropFraction = {1.0f, 1.0f};
    }
    bool centerOrRandomCrop = false; // true for random crop
    std::shared_ptr<ImageDataset::ImageConfig> datasetConfig(ImageDataset::ImageConfig::create(converImagesToFormat, resizeHeight, resizeWidth, scales, means,cropFraction, centerOrRandomCrop));
    bool readAllImagesToMemory = false;
    auto trainDataset = ImageDataset::create(trainImagesFolder, trainImagesTxt, datasetConfig.get(), readAllImagesToMemory);
    auto testDataset = ImageDataset::create(testImagesFolder, testImagesTxt, datasetConfig.get(), readAllImagesToMemory);

    const int trainBatchSize = 32;
    const int trainNumWorkers = 4;
    const int testBatchSize = 10;
    const int testNumWorkers = 0;

    auto trainDataLoader = trainDataset.createLoader(trainBatchSize, true, true, trainNumWorkers);
    auto testDataLoader = testDataset.createLoader(testBatchSize, true, false, testNumWorkers);

    const int trainIterations = trainDataLoader->iterNumber();
    const int testIterations = testDataLoader->iterNumber();

    // const int usedSize = 1000;
    // const int testIterations = usedSize / testBatchSize;

    for (int epoch = 0; epoch < 50; ++epoch) {
        model->clearCache();
        {
            AUTOTIME;
            trainDataLoader->reset();
            model->setIsTraining(true);
            for (int i = 0; i < trainIterations; i++) {
                AUTOTIME;
                auto trainData  = trainDataLoader->next();
                auto example    = trainData[0];
                // Compute One-Hot
                auto newTarget = _OneHot(_Cast<int32_t>(_Squeeze(example.second[0] + _Scalar<int32_t>(addToLabel), {})),
                                  _Scalar<int>(numClasses), _Scalar<float>(1.0f),
                                         _Scalar<float>(0.0f));
                auto predict = _Convert( model->forward(_Convert(example.first[0], NC4HW4)), NCHW);
                auto loss    = _CrossEntropy(predict, newTarget);
                float rate   = LrScheduler::inv(0.0001, solver->currentStep(), 0.0001, 0.75);
                solver->setLearningRate(rate);
                if (solver->currentStep() % 10 == 0) {
                    std::cout << "train iteration: " << solver->currentStep();
                    std::cout << " loss: " << loss->readMap<float>()[0];
                    std::cout << " lr: " << rate << std::endl;
                }
                solver->step(loss);
                exe->gc(Executor::FULL);
            }
        }

        int correct = 0;
        int sampleCount = 0;
        testDataLoader->reset();
        model->setIsTraining(false);
        exe->gc(Executor::PART);

        AUTOTIME;
        for (int i = 0; i < testIterations; i++) {
            auto data       = testDataLoader->next();
            auto example    = data[0];
            auto predict    = model->forward(_Convert(example.first[0], NC4HW4));
            predict         = _ArgMax(predict, 1); // (N, numClasses) --> (N)
            auto label = _Squeeze(example.second[0]) + _Scalar<int32_t>(addToLabel);
            sampleCount += label->getInfo()->size;
            auto accu       = _Cast<int32_t>(_Equal(predict, label).sum({}));
            correct += accu->readMap<int32_t>()[0];

            if ((i + 1) % 10 == 0) {
                std::cout << "test iteration: " << (i + 1) << " ";
                std::cout << "acc: " << correct << "/" << sampleCount << " = " << float(correct) / sampleCount * 100 << "%";
                std::cout << std::endl;
            }
        }
        auto accu = (float)correct / testDataLoader->size();
        // auto accu = (float)correct / usedSize;
        std::cout << "epoch: " << epoch << "  accuracy: " << accu << std::endl;

        {
            auto forwardInput = _Input({1, 3, resizeHeight, resizeWidth}, NC4HW4);
            forwardInput->setName("data");
            auto predict = model->forward(forwardInput);
            Transformer::turnModelToInfer()->onExecute({predict});
            predict->setName("prob");
            std::string fileName = "temp.mobilenetv2.mnn";
            Variable::save({predict}, fileName.c_str());
            ConvertToFullQuant::convert(fileName);
        }
    }
}